Automating with AI — what is the hold-up?

Discussion starter…

I’ve been thinking about this for a little while now: what is the hold-up when trying to automate with AI? In a lot of cases it turns out to be the vendor or product we’re using — sometimes it’s the risk level of the process, sometimes just current priorities. The vendor piece is becoming the most significant show-stopper, and if vendors don’t adapt, we will have to.

Take a recent example of using VendorGiddyUp (a popular SFTP vendor product, masked) to automate a process with AI. Some simple processes might not save much, but for certain projects this is a huge win.

I just completed an entire VendorGiddyUp project with zero code on my part, completely prompt-driven — with some gotchas I’ll explain. I’ve already done the same with Azure Data Factory. I’ve been seeing where this is going for a while. Below is a complete plan to go from fully manual processes to fully agentic as milestones are met. I’m not saying we should allow a fully agentic process given the risk, but you can get pretty close with one or two minor human-in-the-loop approvals.

This is the future: AI drives the logic, and the system engineer or SME reviews the result before committing. To make this possible we need to understand our vendors’ roadmaps — or pivot for our own productivity.

Here’s an AI-automation-driven step plan for VendorGiddyUp. As of this writing we cannot get past step 2 to my knowledge (further validation warranted).

VendorGiddyUp full-automation progression milestones

Example: file sharing both directions — SFTP→SQL and SQL→SFTP.

1. Fully manual (current state)

• Vendor emails to set up SFTP; Ops/vendor share keys.
• Connections created by Ops in the VendorGiddyUp UI.
• File-format expectations communicated via email.
• Project created by DMT in Project Designer.
• Execute project / validate / update project until working.
• Schedule execution.
• Notification to parties when completed.

Change to process: GitHub repo template with VendorGiddyUp documentation (same as version used).

2. Partial automation — AI-assisted project creation ✅ Proven

• Send template emails to set up SFTP (reduces back-and-forth).
• Connections created by Ops in the VendorGiddyUp UI.
• File-format expectations (captured in template).
• GitHub repo template with VS Code AI chat to create project XML (no human coding needed), paste XML into VendorGiddyUp.
• Execute project / validate and use AI chat to iterate until working.
• Schedule execution.
• Notification to parties when completed.

→ Continued success after X attempts, then advance.

Change to process: leverage the VendorGiddyUp API and scripts for connection setup.

3. Additional automation — API-driven connections ⚠️ Needs validation

• Send template emails to set up SFTP.
• Connections created by API call and scripts (HITL approval).
• ⚠️ VendorGiddyUp does not have a documented REST API for resource/connection creation. Options to validate: CLI tooling, admin API, or pre-pooled reusable resources. Action: confirm with vendor support.
• File-format expectations (in template).
• GitHub repo template with VS Code AI chat to create project XML, paste into VendorGiddyUp.
• Execute project / validate and use AI chat until working.
• Schedule execution (✅ can be embedded in project XML).
• Notification to parties when completed.

→ Continued success after X attempts, then advance.

Change to process: create a website for template intake (login required) and introduce an AI agentic process.

4. Additional automation — agentic deployment ✅/⚠️ Mixed

• Send website link to set up SFTP (no email needed).
• Connections created by API call and scripts (HITL approval) — ⚠️ same dependency as #3.
• File-format expectations (in template).
• GitHub repo template with AI agent to deploy solution.
• ✅ Project execution confirmed possible via CLI and REST API POST with a write-capable account.
• Execute project / validate and correct via AI agent (✅ job logs available via API).
• Schedule execution by AI agent (✅ triggers can be defined in project XML).
• AI notification to parties when completed.

→ Continued success after X attempts, then advance.

Change to process: remove HITL approval and create a fully agentic process (maybe not possible unless risk is acceptable).

5. Full agentic automation 🔴 Aspirational — risk-dependent

• Send website link to set up SFTP (no email needed).
• Connections created by API/scripts (no HITL approval; extra security measures).
• 🔴 Blocked unless the connection-creation gap is resolved and the organization accepts the risk of automated credential / connection provisioning.
• File-format expectations (in template).
• AI agent deploys solution, executes, validates, corrects, schedules, and notifies.
• No human in the loop — requires: audit trail, rollback capability, and security review.

The takeaway

The model side is ready. The orchestration side is ready. The blocker is almost always whether the vendor exposes the surfaces an agent needs — and whether your organization is ready to trust an agent to use them. If your vendor isn’t moving, you have two choices: wait, or pivot to something that gets out of your way.